Tensor Voting: A Perceptual Organization Approach to Computer Vision and Machine Learning

1

Tensor Voting: A Perceptual Organization

Approach to Computer Vision and Machine Learning

Philippos MordohaiUniversity of North Carolina, Chapel Hill

http://cs.unc.edu/~mordohai

2

Short Biography

• Ph.D. from University of Southern California with Gérard Medioni (2000-2005)– Perceptual organization– Binocular and multiple-view stereo– Feature inference from images– Figure completion– Manifold learning– Dimensionality estimation– Function approximation

• Postdoctoral researcher at University of North Carolina with Marc Pollefeys (2005-present)– Real-time video-based reconstruction of urban

environments– Multiple-view reconstruction– Temporally consistent video-based reconstruction

3

Audience

• Academia or industry?• Background:

– Perceptual organization– Image processing– Image segmentation– Human perception– 3-D computer vision– Machine learning

• Have you had exposure to tensor voting before?

4

Objectives

• Unified framework to address wide range of problems as perceptual organization

• Applications:– Computer vision problems– Instance-based learning

5

Overview

• Introduction• Tensor Voting• Stereo Reconstruction• Tensor Voting in N-D• Machine Learning• Boundary Inference • Figure Completion• More Tensor Voting Research• Conclusions

6

Motivation

• Computer vision problems are often inverse– Ill-posed– Computationally expensive– Severely corrupted by noise

• Many of them can be posed as perceptual grouping of primitives– Solutions form perceptually salient non-

accidental structures (e.g. surfaces in stereo)– Only input/output modules need to be adjusted

in most cases

7

Motivation

• Develop an approach that is:– General– Data-driven

• Axiom: the whole is greater than the sum of the parts

• Employ Gestalt principles of proximity and good continuation to infer salient structures from data

8

Gestalt Principles

• Proximity

• Similarity

• Good continuation

• Closure

• Common fate

• Simplicity

9

Structural Saliency• Property of structures to stand out due to

proximity and good continuation

• Local responses are not enough• Need aggregation of supportThe smoothness constraint: applies almost

everywhere

Input Human observerEdge detector

10

Integrated Descriptions

• Different types of structures interact – Junctions are intersections of curves that

do not exist in isolation

• Structures have boundaries

11

Desired Properties

• Local, data-driven descriptions– More general, model-free solutions– Local changes affect descriptions locally– Global optimization often requires simplifying

assumptions (NP-complete problems)

• Able to represent all structure types and their interactions

• Able to process large amounts of data

• Robust to noise

12

Beyond 2- and 3-D• Gestalt principles can be applied in any dimension• Coherent data form smooth, salient structures• Positions, velocities and motor commands form

manifolds in N-D

Vijaykumar et al. 2002

13

Perceptual Organization Approaches

• Symbolic methods

• Clustering

• Local interactions

• Inspired by human visual system

14

Symbolic Methods• Operate on symbols not signals• Marr (1982): hierarchical grouping of symbols

– Primal sketch– 2 ½-D sketch– 3-D model

• Lowe (1985): 3-D object recognition based on grouping of edgels– Gestalt principles– Viewpoint invariance– Low likelihood of accidental alignment

• Saund (2003): perceptually closed paths in sketches and drawings– Loosely convex, mostly closed

15

Symbolic Methods

• Mohan and Nevatia (1992): bottom-up hierarchical with increasing levels of abstraction– 3-D scene descriptions from collations of features

• Dolan and Riseman (1992): hierarchical curvilinear structure inference

• Jacobs (1996): salient convex groups as potential object outlines– Convexity, proximity,

contrast

16

Clustering• Jain and Dubes (1988): textbook• Shi and Malik (2000): normalized cuts on

graph– Edges encode affinity between nodes

• Boykov et al (2001): α-expansion algorithm on labeling graph to minimize objective function– Single-node data terms– Pair-wise regularization

terms

17

Methods based on Local Interactions

• Shashua and Ullman (1988): structural saliency due to length and smoothness of curvature of curves going through each token

• Parent and Zucker (1989): trace points, tangents and curvature from noisy data

• Sander and Zucker (1990): 3-D extension• Guy and Medioni (1996, 1997):

predecessor of tensor voting– Voting fields– Tensor analysis for feature inference– Unified detection of surfaces,

curves and junctions

18

Inspired by Human Visual System

• Grossberg and Mingolla (1985), Grossberg and Todorovic (1988): Boundary Contour System and Feature Contour System– BCS: boundary detection, competition and cooperation,

includes cells that respond to “end-stopping”– FCS: surface diffusion mechanism limited by BCS

boundaries• Heitger and von der Heydt (1993): curvels

grouped into contours via convolution with orientation-selective kernels– Responses decay with distance, difference in

orientation– Detectors for endpoints, T-junctions and corners– Orthogonal and parallel grouping– Explain occlusion, illusory contours consistently with

perception

19

Inspired by Human Visual System

• Williams and Jacobs (1997): stochastic completion fields– Probabilistic model based on

random walks in image lattice• Li (1998): contour integration

with excitatory and inhibitory fields

• Yen and Finkel (1998): voting-based approach – Votes along tangent of osculating

circle at receiver, attenuate with distance

20

Differences with our Approach

• Infer all structure types simultaneously and allow interaction between them

• Can begin with oriented or unoriented inputs (or both)

• No prior model• No objective/cost function• Solution emerges from data

21

Overview

• Introduction• Tensor Voting• Stereo Reconstruction• Tensor Voting in N-D• Machine Learning• Boundary Inference • Figure Completion• Conclusions

22

The Original Tensor Voting Framework

• Perceptual organization of generic tokens [Medioni, Lee, Tang 2000]

• Data representation: second order, symmetric, nonnegative definite tensors

• Information propagation: tensor voting• Infers saliency values and preferred

orientation for each type of structure

23

Second Order Tensors• Symmetric, non-negative definite• Equivalent to:

– Ellipse in 2-D or ellipsoid in 3-D– 2x2 or 3x3 matrix

• Properties that can be encoded:– shape: orientation certainty

– size: feature saliency

24

Second Order Tensors in 2-D

• 2×2 Matrix or Ellipse can be decomposed:– Stick component– Ball component

+=

00

02

22

22

22

222 b

aa

aa

aa

aba

)()( 221121121

222111

TTT

TT

eeeeee

eeeeT

25

Representation with Tensors

26

Tensor Voting

?

27

Saliency Decay Function

• Votes attenuate with length of smooth path and curvature

• Stored in pre-computed voting fields

x

y

P

s

C

O

l

l

ls

sin2sin

28

Fundamental Stick Voting Field

29

Fundamental Stick Voting Field

All other fields in any N-D space are generated from the Fundamental Stick Field:

– Ball Field in 2-D– Stick, Plate and Ball Field in 3-D– Stick, …, Ball Field in N-D

30

2-D Ball Field

Ball field computed by integrating the contributions of rotating stick

dPP )()( SB

S(P) B(P)

PP

31

2-D Voting Fields

votes with

votes with

votes with +

Each input site propagates its information in a neighborhood

32

Voting

• Voting from a ball tensor is isotropic – Function of distance only

• The stick voting field is aligned with the orientation of the stick tensor

O

P

33

Scale of Voting

• The Scale of Voting is the single critical parameter in the framework

• Essentially defines size of voting neighborhood– Gaussian decay has infinite extend, but

it is cropped to where votes remain meaningful (e.g. 1% of voter saliency)

34

Scale of Voting

• The Scale is a measure of the degree of smoothness

• Smaller scales correspond to small voting neighborhoods, fewer votes– Preserve details– More susceptible to outlier corruption

• Larger scales correspond to large voting neighborhoods, more votes– Bridge gaps– Smooth perturbations– Robust to noise

35

Vote Accumulation

Each site accumulates second order votes by tensor addition:

Results of accumulation are usually generic tensors

36

Vote Analysis

• λ1- λ2> λ2: stick saliency is larger than ball saliency. Likely on curve.

• λ1≈λ2>0: ball saliency larger than stick saliency. Likely junction or region.

• λ1≈λ2 ≈ 0: Low saliency. Outlier.

)()( 221121121

222111

TTT

TT

eeeeee

eeeeT

37

Junction or Region Inlier?

Input

Ball saliencymap

38

Results in 2-D

39

Sensitivity to Scale

AB

Input

= 50

= 500

= 5000Curve saliency as a function of scaleBlue: curve saliency at ARed: curve saliency at B

Input: 166 un-oriented inliers, 300 outliersDimensions: 960x720Scale ∈ [50, 5000] Voting neighborhood ∈ [12, 114]

40


Circle with radius 100 (unoriented tokens)

As more information is accumulated, the tokens better approximate the circle

41


Square 20x20 (unoriented tokens)

As scale increases to unreasonable levels (>1000)corners get rounded

Junctions are detected and excluded

42

Structural Saliency Estimation• Data from [Williams and Thornber IJCV 1999]• Foreground objects (N edgels) on background clutter• Detect N most salient edgels

43

Structural Saliency Estimation

• SNR: ratio of foreground edgels to background edgels

• FPR: false positive rate for foreground detection

• Our results outperform all methods evaluated in [Williams and Thornber IJCV 1999]

SNR 25 20 15 10 5

FPR 10.0%

12.4%

18.4%

35.8%

64.3%

44

Structure Types in 3-D

The input may consist of

point/junction curvel

surfel

45

3-D Second Order Tensors

• Encode normal orientation in tensor

• Surfel: 1 normal “stick” tensor

• Curvel: 2 normals “plate” tensor

• Point/junction: 3 normals “ball” tensor

46

Representation in 3-D

47

3-D Tensor Analysis

• Surface saliency: λ1- λ2 normal: e1

• Curve saliency: λ2- λ3 normals: e1 and e2

• Junction saliency: λ3

)())(()( 33221132211321121

333222111

TTTTTT

TTT

eeeeeeeeeeee

eeeeee

T

48

3-D Voting Fields

• 2-D stick field is a cut of the 3-D field containing the voter

• Plate and ball fields derived by integrating contributions of rotating stick voter– Stick spans disk and sphere

respectively

49

Vote Analysis in 3-D

• λ1- λ2>λ2- λ3 and λ1- λ2>λ3: stick saliency is maximum. Likely surface.

• λ2- λ3>λ1- λ2 and λ2- λ3>λ3 : plate saliency is maximum. Likely curve or surface intersection

• λ3>λ1- λ2 and λ3> λ2- λ3 : ball saliency is maximum. Likely junction

• λ1≈λ2 ≈λ3 ≈ 0: Low saliency. Outlier.

)())(()( 33221132211321121

333222111

TTTTTT

TTT

eeeeeeeeeeee

eeeeee

T

50

Results in 3-D

Noisy input

Surface intersections Dense surfaces

Surface inliers

51

Overview


52

Approach for Stereo

• Problem can be posed as perceptual organization in 3-D – Correct pixel matches should form

smooth, salient surfaces in 3-D– 3-D surfaces should dictate pixel

correspondences• Infer matches and surfaces by

tensor voting• Use monocular cues to

complement binocular matches [Mordohai and Medioni, ECCV 2004 and PAMI 2006]

53

Challenges

• Major difficulties in stereo: – occlusion – lack of texture

• Local matching is not always reliable:– False matches can have high scores

54

Algorithm Overview

• Initial matching• Detection of

correct matches • Surface grouping

and refinement• Disparity

estimation for unmatched pixels

55

Initial Matching

• Each matching technique has different strengths

• Use multiple techniques and both images as reference:– 5×5 normalized cross correlation (NCC)

window– 5×5 shiftable NCC window– 25×25 NCC window for pixels with very low

color variance– 7×7 symmetric interval matching window

with truncated cost function

Note: small windows produce random (not systematic) errors (reduced foreground fattening)

56

Symmetric Interval Matching

• Upsample scanlines• Represented the color

of pixel (x,y) as the interval (x-½, y) to (x+½, y)

• Dissimilarity measure: distance between intervals

• Truncated to increase performance at discontinuities

d

d=0

57

Candidate Matches

• Compute sub-pixel estimates (parabolic fit)

• Keep all good matches (peaks of NCC)• Drop scores

– Depend on texture properties– Hard to combine across methods

58

Candidate Matches

59

Surfaces from Unoriented Data

• Voting is pair-wise• Two unoriented

tokens define a path and the voter casts a vote (normal spans plane)

• Accumulation of votes with a common axis results in a salient surface normal

A

C

B

60

Detection of Correct Matches

• Tensor voting performed in 3-D

• Saliency used as criterion to disambiguate matches instead of aggregated matching cost or global energy

• Visibility constraint enforced along rays with respect to surface saliency

A

CD

B

61

Uniqueness vs. Visibility

• Uniqueness constraint: One-to-one pixel correspondence– Exact only for fronto-parallel surfaces

• Visibility constraint : M-to-N pixel correspondences– [Ogale and Aloimonos 2004][Sun et al.

2005]– One match per ray of each camera

62

Surface Grouping

• Image segmentation has been shown to help stereo – Not an easier problem

• Instead, group candidate matches in 3-D based on geometric properties– Pick most salient candidate matches as

seeds– Grow surfaces

• Represent surfaces as local collections of colors

63

Nonparametric Color Model• Each match has been

assigned to a surface now– Pixel on match’s ray takes

same label

• No difficulties caused by:– Adjacent image regions with

similar color properties – Surfaces with varying color

distribution

• A GMM would be more complicated and not exact

64

Surface Refinement• Re-visit pixels and verify label assignments• Find all neighboring surfaces including current

assignment s• Compute Ri(x0,y0) for all surfaces

• Ratio of pixels of surface i within neighborhood N similar in color to IL(x0,y0) over all pixels in N labeled as i

• Remove match if Rs(x0,y0) is not maximum Set of reliable matches (reduced foreground fattening)

65

Surface Refinement Results

144808 matches4278 errors


66




67




68


69

Disparity Hypotheses for Unmatched Pixels

• Check color consistency with nearby layers (on both images if not occluded)

• Generate hypotheses for membership in layers with similar color properties– Find disparity range from

neighbors (and extend)– Allow occluded hypotheses– Do not allow occlusion of

reliable matches

• Progressively increase color similarity tolerance and scale

70

Disparities for Unmatched Pixels

• Vote from neighbors of same surface

• Keep most salient– Update occlusion

information

d

x

y

Disparity Map Error Map

71

Results: Tsukuba

Left image Ground truth

Error MapDisparity Map

72

Results: Venus



73

Results: Cones



74

Results: Teddy



75

Quantitative Results

76

Aerial Images

77

Summary of Approach to Stereo

• Binocular and monocular cues are combined• Novel initial matching framework• No image segmentation • Occluding surfaces do not over-extend

because of color consistency requirement• Textureless surfaces are inferred based on

surface smoothness– When initial matching fails

78

Multiple-View Stereo

• Approach to dense multiple-view stereo– Multiple views: more than two– Dense: attempt to reconstruct all pixels

• Process all data simultaneously– Do not rely on binocular results– Only binocular step: detection of potential pixel

correspondences

• Correct matches form coherent salient surfaces– Infer them by Tensor Voting

[Mordohai and Medioni, 3DPVT 2004]

79

Desired Properties

• General camera placement – As long as camera pairs are close enough for

automatic pixel matching• No privileged images• Features required to appear in no more

than two images• Reconstruct background

– Do not discard• Simultaneous processing

– Do not merge binocular results

80

Input Images

Captured at the CMU dome for the Virtualized Reality project

81

Limitations of Binocular Processing

Errors due to:– Occlusion– Lack of texture

Matching candidates in disparity space Disparity map after binocular processing

82

Limitations of Binocular Processing

When matching from many image pairs, candidates are combined:

– Lessens effects of depth discontinuities• Occluded surfaces are revealed

– Salient surfaces are reinforced by correct matches from multiple pairs• Noise is not

83

Candidate Matches

10 image pairs, 1.1 million pointsTensor voting takes 44 min. 30 sec. (March 2004)

84

Results on “Meditation” Set

85

Results on “Meditation” Set

86

Results on “Baseball” Dataset

87

Overview


88

Tensor Representation in N-D• Non-accidental alignment, proximity, good

continuation apply in N-D– Robot arm moving from point to point forms 1-D

trajectory (manifold) in N-D space• Noise robustness and ability to represent all

structure types also desirable

• Tensor construction: – eigenvectors of normal space associated with

non-zero eigenvalues– eigenvectors of tangent space associated with

non-zero eigenvalues

[Mordohai, PhD Thesis 2005]

89

Limitations of Voting Fields

• Hard to generalize to N dimensions• Requirements: N N-D fields

– k samples per axis: O(NkN) storage requirements

– Nth order integration to compute each sample

S(P) B(P)P

P

dPP )()( SB

90

Efficient N-D implementation• Drop uncertainty from vote computation• Cast votes directly without integration

– Votes from stick tensors are computed in 2-D subspace regardless of N

– Ball tensors cast votes that support straight lines from voter to receiver

91

Efficient N-D implementation• Simple geometric solution for arbitrary tensors• Observation: curvature only needed for saliency

computation when θ not zero

• vn projection of vector AB on normal space of voter

• Define basis for voter that includes vn

– 1 vote computation that requires curvature– At most N-2 vote computations that are scaled stick

tensors parallel to voters

92

Vote Analysis

• Tensor decomposition:

• Dimensionality estimate: d with max{λd- λd+1}

• Orientation estimate: normal subspace spanned by d eigenvectors corresponding to d largest eigenvalues

d

TddN

TTT

TNNN

TT

eeeeeeee

eeeeee

...))(()(

...

2211321121

222111T

93

Comparison with Original Implementation

• Tested on 3-D data• Saliency maps qualitatively

equivalent

Old

New

Surface saliencyz=120

Surface saliencyz=120

Curve saliencyz=0

94


• Surface orientation estimation – Inputs encoded as ball tensors

• Slightly in favor of new implementation– Pre-computed voting fields use

interpolation

95


• Noisy data

5:1 outlier to inlier ratio

96

Overview


97

Instance-based Learning

• Learn from observations in continuous domain

• Observations are N-D vectors• Estimate:

– Intrinsic dimensionality– Orientation of manifold through each

observation– Generate new samples on manifold

98

Approach

• Vote as before in N-D space• Intrinsic dimensionality found as

maximum gap of eigenvalues• Do not perform dimensionality reduction• Applicable to:

– Data of varying dimensionality– Manifolds with intrinsic curvature (distorted

during unfolding)– “Unfoldable” manifolds (spheres, hyper-

spheres)– Intersecting manifolds

99

Approach

• Manifold orientation from eigenvectors

• “Eager learning” since all inputs are processed and queries do not affect estimates– “Lazy” alternative: collect votes on

query points– Requires stationary data distribution

100

Dimensionality Estimation

“Swiss Roll”

20,000 points on 2-D manifold in 3-D

[Mordohai and Medioni, IJCAI 2005]

101

Dimensionality Estimation• Synthetic data• Randomly sample input variables

(intrinsic dimensionality)• Map them to higher dimensional

vector using linear and quadratic functions

• Add noisy dimensions• Global rotation

102

Dimensionality Estimation• Point-wise dimensionality estimates• No global operations

Input in 4-D

1-D

3-D2-D

103

Manifold Learning

• Input: instances (points in N-D space)• Try to infer local structure (manifold)

assuming coherence of underlying mechanism that generates instances

• Tensor voting provides:– Dimensionality– Orientation

104

Orientation Estimation

• Functions proposed by [Wang et al. 2004]• Challenge: non-uniform sampling

Mean error: 0.40 degrees for 152 samples

Mean error: 2.02 degrees for 180 samples0.57 degrees for 360 samples

)

105

Manifold Distance Measurement

• Do not reduce dimensionality

• Start from point on manifold

• Take small step along desired orientation on tangent space

• Generate new point and collect votes

• Repeat until convergence

AA1

B

p

106

Distance Measurement: Test Data

Error in orientation estimation: 0.11o 0.26o

• Test data: spherical and cylindrical sections– Almost uniformly sampled– Ground truth distances between points are known

• Goal: compare against algorithms that preserve pair-wise properties of being far away or close

107

Experimental Setup

• Comparison of five leading manifold learning algorithms and our approach in distance measurement

• Randomly select pairs of points on the manifold, measure their distance in embedded space and compare with ground truth– Apply uniform scaling for algorithms where original

distance metric is not preserved

108

Distance Measurement with Outliers

109

Traveling on Manifolds

50-D30-D

110

Function Approximation• Problem: given point Ai in input space predict output value(s)• Find neighbor Bi with known output• Starting from B in joint input-output space, interpolate until A

is reached– A projects on input space within ε of Ai

111

Synthetic Data

• Sample 1681 points from

proposed by Schaal and Atkenson, 1998• Perform tensor voting and generate new points

– On original noise-free data– On data with 8405 outliers– On data with 8405 outliers and Gaussian perturbation

of the inliers (σ=0.1)– Same data embedded in 60-D

112

Noisy input

New points

Synthetic Data

113

New points for data with outliersand perturbation in 60-D

NMSE: MSE normalized by variance of noise free input data

Synthetic Data

114

Function approximation on datasets from:• University of California at Irvine Machine

Learning Repository• DELVE archive

• Rescale data (manually) so that dimensions become comparable (variance 1:10 instead of original 1000:1)

• Split randomly into training and test sets according to literature– Repeat several times

Real Data

115

Comparable with recently published results using Bayesian Committee Machine, Gaussian Process Regression, Support Vector Regression etc.

Results on Real Data

Dataset Dim. Training Test Mean Error

Abalone 9 3000 1177 1.63

Boston Housing 13 481 25 1.27

Computer Activity 22 2000 6192 1.97

116

Advantages over State of the Art

• Broader domain for manifold learning– Manifolds with intrinsic curvature (cannot be

unfolded)– Open and closed manifolds (hyper-spheres)– Intersecting manifolds– Data with varying dimensionality

• No global computations O(NM logM)• Noise Robustness

• Limitations:– Need more data– “Regular” distribution

117

Overview


118

Boundary Inference• Second-order tensors can

represent second order-discontinuities – Discontinuous orientation (A)

• But not first-order discontinuities– Discontinuous structure (C)

Tensor with dominant plate component(orthogonal to surface intersection)

A:A

B

C How to discriminate B from C?

119

Boundary Inference: First Order Properties

• Representation augmented with Polarity Vectors

• Sensitive to direction from which votes are received

• Boundaries have all their neighbors on the same side of the half-space

x

y

P

s

C

O

l

Second order vote

First order vote

[Tong, Tang, Mordohai and Medioni, PAMI 2004]

120

Illustration of Polarity

Input

Curve Saliency

Polarity

121

Illustration of First Order Voting

Tensor Voting with first order properties

122

Illustration of Region Inference

123


124

Results in 2-D

Gray: curve inliersBlack: curve endpointsSquares: junctions

Input

125

Results in 2-D

Input Curves and endpoints onlyCurves, endpoints, regions and region boundaries

126


127

Results in 3-D

Input Surfaces - Surface Boundaries – Surface IntersectionsCurves – Endpoints - Junctions

128

Results in 3-D

Noisy input Dense surface boundary(after marching cubes)

129

Results in 3-D

Input (600k unoriented points) Input with 600k outliers

Output with 1.2M outliers Output with 2.4M outliers

130

Overview


131

Figure Completion

Amodal completion

Modal completion Layered interpretation

132

Motivation

• Approach for modal and amodal

completion

• Automatic selection between them

• Explanation of challenging visual

stimuli consistent with human

visual system

[Mordohai and Medioni, POCV 2004]

133

Keypoint Detection

• Input binary images• Infer junctions, curves, endpoints,

regions and boundaries• Look for completions supported by

endpoints, L and T-junctions • W, X and Y-junctions do not

support completion by themselves

134

Support for Figure Completion

• Amodal:– Along the tangent of endpoints– Along the stem of T-junctions

• Modal:– Orthogonal to endpoints– Along the bar of T-junctions– Along either edge of L-junctions

135

Voting for Completion

• At least two keypoints of appropriate type needed

• Possible cases:– No possible continuation– Possible amodal completion (parallel

field)– Possible modal completion

(orthogonal field)– If both possibilities available, modal

completion is perceived as occluding amodal one

136

Results: Modal Completion

Input Curves and endpoints Curve saliency Output

137

The Koffka Cross

Input

Curve saliency

OutputJunction saliency

138

The Koffka Cross

Input

Curve saliency


Note: maximum junction saliency here is 90% of maximum curve saliency, but only 10% in the previous case

139

Koffka Cross: Amodal Completion

Curve saliency


Amodal completion(occluded)

140

Koffka Cross: Both Types of Completion

Input Output

141

The Poggendorf Illusion

Input OutputCurve saliency

142

The Poggendorf Illusion

Input OutputCurve saliency

143

Discussion• Current approach:

– Implements modal and amodal completion and automatically selects appropriate type

– Interprets correctly complex perceptual phenomena

• More work needed on:– L-junctions which offer two alternatives– Inference of hierarchical descriptions

144

Overview


145

More Tensor Voting Research

• Curvature estimation• Visual motion analysis• Epipolar geometry estimation for

non-static scenes• Texture synthesis

146

Curvature Estimation

• Challenging for noisy, irregular point cloud

• Three passes– Estimate surface normals– Compute subvoxel updates for

positions of points– Compute curvature by collecting votes

from 8 directions (45o apart)

• Infer dense surface

[Tong and Tang, PAMI 2005]

147

Visual Motion Analysis

• Grouping exclusively based on motion cues [Nicolescu and Medioni, ICPR 2002]

• Motion-segmentation on real images– Accurate object boundaries

148

Visual Motion Analysis

• Potential matches represented by 4-D tensors (x, y, vx, vy)

• Desired motion layers have maximum λ2-λ3

• Results on challenging, non-rigid datasets

149

Motion Segmentation

• Group candidate matches in “surfaces” as in stereo

• Boundaries may be inaccurate– Use intensity cues (edges) from original images– Infer most salient curves in vicinity or object boundaries

• Parallel to initial boundaries

Blue: initial boundaries Red: refined boundaries

[Nicolescu and Medioni, PAMI 2005]

150

Epipolar Geometry Estimation• Epipolar geometry defines a 4-D point cone

in the joint image space [Anandan, ECCV 2000]

• Vote in 4-D to detect points on cone or cones– Each cone corresponds to an epipolar geometry

(u1,v1) (u2,v2)

(u1,v1,u2,v2)

Image 1 Image 2

4-D point cone

Joint image space[Tong, Tang and Medioni, PAMI 2004]

151

Epipolar Geometry Estimation

152

Texture Synthesis

• Given image with user-specified target region• Segment • Connect curves across target region• Synthesize texture via N-D tensor voting

[Jia and Tang, PAMI 2004]

153

Texture Synthesis Results

154

Texture Synthesis Results

155

Overview


156

Conclusions

• General framework for perceptual organization

• Unified and rich representation for all types of structure, boundaries and intersections

• Model-free

• Applications in several domains

157

Future Work

• Integrated feature detection in real images [Förstner94, Köthe03]– Integrated inference of all feature types (step

and roof edges, all junction types)

• Decision making strategies for interpretation as in [Saund 2003] for binary images and sketches– Symmetry, length, parallelism– Good continuation vs. maximally turning

paths

• Hierarchical descriptions

158

Future Work

• Rich monocular descriptions in terms of:– Shape (invariant as possible)– Occlusion, depth ordering, completion

• Applications– Reconstruction– Recognition

159

Future Work

• Extend manifold learning work– Classification– Complex systems such as forward

and inverse kinematics– Data mining

• Learn from imperfect descriptions– Symbolic not signal-based– Overcome limitations of image-as-

vector-of-pixels representation

Documents

Tensor Voting: A Perceptual Organization Approach to Computer Vision and Machine Learning